时间序列复习
密码 2022年6月11日-6月13日
文章目录
Chapter 7 Cointegration and Error Correction Models
Long-run Equilibrium
-
The formal analysis begins by considering a set of economic variables in long-run equilibrium when
β 0 , 1 z 1 t + β 0 , 2 z 2 t + ⋯ + β 0 , k z k t = 0. \beta_{0,1} z_{1 t}+\beta_{0,2} z_{2 t}+\cdots+\beta_{0, k} z_{k t}=0 . β0,1z1t+β0,2z2t+⋯+β0,kzkt=0. -
Let β 0 = ( β 0 , 1 , ⋯ , β 0 , k ) ′ \beta_{0}=\left(\beta_{0,1}, \cdots, \beta_{0, k}\right)^{\prime} β0=(β0,1,⋯,β0,k)′ and z t = ( z 1 t , ⋯ , z k t ) ′ ∼ I ( 1 ) z_{t}=\left(z_{1 t}, \cdots, z_{k t}\right)^{\prime} \sim I(1) zt=(z1t,⋯,zkt)′∼I(1).
-
The system is in long-run equilibrium when β 0 ′ z t = 0 \beta_{0}^{\prime} z_{t}=0 β0′zt=0.
-
**The deviation from long-run equilibrium- called the equilibrium error-**is e t e_{t} et, so that e t = β 0 ′ z t e_{t}=\beta_{0}^{\prime} z_{t} et=β0′zt.
-
If the equilibrium is meaningful, it must be the case that the equilibrium error process e t e_{t} et is stationary.
-
The cointegrating vector β 0 \beta_{0} β0 is not unique since for any scalar c, c β 0 ′ z t = β ∗ ′ z t ∼ I ( 0 ) c \beta_{0}^{\prime} z_{t}=\beta^{*^{\prime}} z_{t} \sim I(0) cβ0′zt=β∗′zt∼I(0).
-
Some normalization is required to uniquely identify β 0 \beta_{0} β0. ==A typical normalization is β 0 = ( 1 , − β 0 , 2 , ⋯ , − β 0 , k ) ′ \beta_{0}=\left(1,-\beta_{0,2}, \cdots,-\beta_{0, k}\right)^{\prime} β0=(1,−β0,2,⋯,−β0,k)′,==so that
β 0 ′ z t = z 1 t − β 0 , 2 z 2 t − ⋯ − β 0 , k z k t = e t ∼ I ( 0 ) . \beta_{0}^{\prime} z_{t}=z_{1 t}-\beta_{0,2} z_{2 t}-\cdots-\beta_{0, k} z_{k t}=e_{t} \sim I(0) . β0′zt=z1t−β0,2z2t−⋯−β0,kzkt=et∼I(0).
协整的空间
- If the ( k × 1 ) (k \times 1) (k×1) vector z t z_{t} zt is cointegrated there may be 1 ≤ m ≤ k − 1 1 \leq m \leq k-1 1≤m≤k−1 linearly independent cointegrating vectors { β 1 , ⋯ , β m } \left\{\beta_{1}, \cdots, \beta_{m}\right\} {β1,⋯,βm}.
- β i ′ z t ∼ I ( 0 ) \beta_{i}^{\prime} z_{t} \sim I(0) βi′zt∼I(0) for i = 1 , ⋯ , m i=1, \cdots, m i=1,⋯,m.
- The ( k × m ) (k \times m) (k×m) matrix β = ( β 1 , ⋯ , β m ) \beta=\left(\beta_{1}, \cdots, \beta_{m}\right) β=(β1,⋯,βm) forms a basis for the space of cointegrating vectors.
- Any linear combination of { β 1 , ⋯ , β m } \left\{\beta_{1}, \cdots, \beta_{m}\right\} {β1,⋯,βm} is also a cointegrating vector.
Bivariate VAR(1) Containing Unit Roots
z
t
=
Φ
0
+
Φ
1
z
t
−
1
+
a
t
,
or
Δ
z
t
=
Φ
0
+
(
Φ
1
−
l
)
z
t
−
1
+
a
t
z_{t}=\Phi_{0}+\Phi_{1} z_{t-1}+a_{t}, \text { or } \Delta z_{t}=\Phi_{0}+\left(\Phi_{1}-l\right) z_{t-1}+a_{t}
zt=Φ0+Φ1zt−1+at, or Δzt=Φ0+(Φ1−l)zt−1+at
The stationarity of
z
t
=
[
y
t
,
x
t
]
′
z_{t}=\left[y_{t}, x_{t}\right]^{\prime}
zt=[yt,xt]′ requires that roots of
det
(
I
−
Φ
1
λ
)
=
0
\operatorname{det}\left(I-\Phi_{1} \lambda\right)=0
det(I−Φ1λ)=0
are all larger than 1 in modulus.
-
z t z_{t} zt has unit roots, therefore unity solves Eq.(3), i.e. det ( I − Φ 1 ) = 0 \operatorname{det}\left(I-\Phi_{1}\right)=0 det(I−Φ1)=0. In other words, Π ≡ Φ 1 − I \Pi \equiv \Phi_{1}-I Π≡Φ1−I is singular.(奇异的)
-
Case 1 : rank ( Π ) = 0 1: \operatorname{rank}(\Pi)=0 1:rank(Π)=0, i.e. Φ 1 = I \Phi_{1}=I Φ1=I.
- It is easy to check that Eq.(3) has two unit roots. No common trend! No cointegration!
- Eq.(96)VAR(1) becomes
Δ z t = Φ 0 + a t , VAR in first differences \Delta z_{t}=\Phi_{0}+a_{t}, \quad \text { VAR in first differences } Δzt=Φ0+at, VAR in first differences
-
Case 2 : rank ( Π ) = 1 \operatorname{rank}(\Pi)=1 rank(Π)=1. Let Φ 1 = [ ϕ 11 ϕ 12 ϕ 21 ϕ 22 ] \Phi_{1}=\left[\begin{array}{cc}\phi_{11} & \phi_{12} \\ \phi_{21} & \phi_{22}\end{array}\right] Φ1=[ϕ11ϕ21ϕ12ϕ22].
-
Let λ 1 , λ 2 \lambda_{1}, \lambda_{2} λ1,λ2 be the roots to Eq.(3): λ 1 = 1 \lambda_{1}=1 λ1=1 and ∣ λ 2 ∣ > 1 \left|\lambda_{2}\right|>1 ∣λ2∣>1.
-
Denote β 0 = ϕ 12 1 − ϕ 11 , α 1 = ϕ 11 − 1 \beta_{0}=\frac{\phi_{12}}{1-\phi_{11}}, \alpha_{1}=\phi_{11}-1 β0=1−ϕ11ϕ12,α1=ϕ11−1 and α 2 = ϕ 21 \alpha_{2}=\phi_{21} α2=ϕ21. Then we have
-
Π = Φ 1 − I = [ α 1 − α 1 β 0 α 2 − α 2 β 0 ] \Pi=\Phi_{1}-I=\left[\begin{array}{ll} \alpha_{1} & -\alpha_{1} \beta_{0} \\ \alpha_{2} & -\alpha_{2} \beta_{0} \end{array}\right] Π=Φ1−I=[α1α2−α1β0−α2β0]
Error-Correction Model
error-correction models:
In Case 2, Eq.(2) becomes
Δ
z
t
=
Φ
0
+
[
α
1
α
2
]
[
y
t
−
1
−
β
0
x
t
−
1
]
+
a
t
,
or
Δ
y
t
=
ϕ
10
+
α
1
[
y
t
−
1
−
β
0
x
t
−
1
]
+
a
1
t
,
Δ
x
t
=
ϕ
20
+
α
2
[
y
t
−
1
−
β
0
x
t
−
1
]
+
a
2
t
,
\begin{aligned} &\Delta z_{t}=\Phi_{0}+\left[\begin{array}{l} \alpha_{1} \\ \alpha_{2} \end{array}\right]\left[y_{t-1}-\beta_{0} x_{t-1}\right]+a_{t}, \\ &\text { or } \begin{array}{l} \Delta y_{t}=\phi_{10}+\alpha_{1}\left[y_{t-1}-\beta_{0} x_{t-1}\right]+a_{1 t}, \\ \Delta x_{t}=\phi_{20}+\alpha_{2}\left[y_{t-1}-\beta_{0} x_{t-1}\right]+a_{2 t}, \end{array} \end{aligned}
Δzt=Φ0+[α1α2][yt−1−β0xt−1]+at, or Δyt=ϕ10+α1[yt−1−β0xt−1]+a1t,Δxt=ϕ20+α2[yt−1−β0xt−1]+a2t,
where in general
α
1
<
0
\alpha_{1}<0
α1<0 and
α
2
>
0
\alpha_{2}>0
α2>0. Eqs. (4) and (5) are called error-correction models.
- In an error-correction model, the short-term dynamics of the variables in the system are influenced by the deviation from equilibrium. 在纠错模型中,系统中变量的短期动态受到平衡偏差的影响
- Granger representation theorem 对于任何I(1)变量集,误差修正和协积分都是等价的表示。
Granger Representation Theorem
If
0
<
rank
(
Π
)
=
m
<
k
0<\operatorname{rank}(\Pi)=m<k
0<rank(Π)=m<k
- Granger representation theorem : Π = α β ′ \Pi=\alpha \beta^{\prime} Π=αβ′, where α \alpha α and β \beta β are k × m k \times m k×m matrices with full column rank.
- β = [ β 1 , ⋯ , β m ] \beta=\left[\beta_{1}, \cdots, \beta_{m}\right] β=[β1,⋯,βm] are the m \mathbf{m} m linearly independent cointegrating vectors. e t = [ e 1 t , ⋯ , e m t ] ′ = β ′ z t ∼ I ( 0 ) e_{t}=\left[e_{1 t}, \cdots, e_{m t}\right]^{\prime}=\beta^{\prime} z_{t} \sim I(0) et=[e1t,⋯,emt]′=β′zt∼I(0) are the m m m equilibrium errors.
- α = [ α 1 , ⋯ , α m ] \alpha=\left[\alpha_{1}, \cdots, \alpha_{m}\right] α=[α1,⋯,αm] collect the adjustment coefficients.
2-variable VAR§ Containing Unit Roots
- Assume each component series contains only one unit root.
- VAR§ : z t = Φ 0 + Φ 1 z t − 1 + ⋯ + Φ p z t − p + a t z_{t}=\Phi_{0}+\Phi_{1} z_{t-1}+\cdots+\Phi_{p} z_{t-p}+a_{t} zt=Φ0+Φ1zt−1+⋯+Φpzt−p+at
- As we did in the univariate case, we can rewrite the model as
Δ z t = η 1 Δ z t − 1 + ⋯ + η p − 1 Δ z t − p + 1 + Φ 0 + a t + [ Φ 1 + Φ 2 + ⋯ + Φ p − l 2 ] z t − 1 ⏟ error correction term , \begin{aligned} \Delta z_{t} &=\eta_{1} \Delta z_{t-1}+\cdots+\eta_{p-1} \Delta z_{t-p+1}+\Phi_{0}+a_{t} \\ &+\underbrace{\left[\Phi_{1}+\Phi_{2}+\cdots+\Phi_{p}-l_{2}\right] z_{t-1}}_{\text {error correction term }}, \end{aligned} Δzt=η1Δzt−1+⋯+ηp−1Δzt−p+1+Φ0+at+error correction term [Φ1+Φ2+⋯+Φp−l2]zt−1,
where η j ≡ − [ Φ j + 1 + Φ j + 2 + ⋯ + Φ p ] \eta_{j} \equiv-\left[\Phi_{j+1}+\Phi_{j+2}+\cdots+\Phi_{p}\right] ηj≡−[Φj+1+Φj+2+⋯+Φp] for j = 1 , 2 , ⋯ , p − 1 j=1,2, \cdots, p-1 j=1,2,⋯,p−1. - The matrix Π ≡ Φ 1 + Φ 2 + ⋯ + Φ p − l 2 \Pi \equiv \Phi_{1}+\Phi_{2}+\cdots+\Phi_{p}-l_{2} Π≡Φ1+Φ2+⋯+Φp−l2 is crucial for identifying the cointegration structure.
Case 1 : No cointegration Π = 0 \Pi=0 Π=0
-
The first difference of z z z follows a stationary VAR (p-1) process
Δ z t = η 1 Δ z t − 1 + ⋯ + η p − 1 Δ z t − p + 1 + Φ 0 + a t . \Delta z_{t}=\eta_{1} \Delta z_{t-1}+\cdots+\eta_{p-1} \Delta z_{t-p+1}+\Phi_{0}+a_{t} . Δzt=η1Δzt−1+⋯+ηp−1Δzt−p+1+Φ0+at.
VAR in first differences! -
We can show that this system has two unit roots. No common trend! No cointegration!
Case 2 : Cointegration
-
Π = Φ 1 + Φ 2 + ⋯ + Φ p − I 2 \Pi=\Phi_{1}+\Phi_{2}+\cdots+\Phi_{p}-I_{2} Π=Φ1+Φ2+⋯+Φp−I2 is singular but nonzero, i.e. rank ( Π ) = 1 \operatorname{rank}(\Pi)=1 rank(Π)=1
-
We have Π = [ α 1 − α 1 β 0 α 2 − α 2 β 0 ] \Pi=\left[\begin{array}{cc}\alpha_{1} & -\alpha_{1} \beta_{0} \\ \alpha_{2} & -\alpha_{2} \beta_{0}\end{array}\right] Π=[α1α2−α1β0−α2β0]. Thus Eq.(6) becomes
Δ z t = Φ 0 + [ α 1 α 2 ] [ y t − 1 − β 0 x t − 1 ] + ∑ j = 1 p − 1 η j Δ z t − j + a t , or \Delta z_{t}=\Phi_{0}+\left[\begin{array}{l} \alpha_{1} \\ \alpha_{2} \end{array}\right]\left[y_{t-1}-\beta_{0} x_{t-1}\right]+\sum_{j=1}^{p-1} \eta_{j} \Delta z_{t-j}+a_{t}, \quad \text { or } Δzt=Φ0+[α1α2][yt−1−β0xt−1]+j=1∑p−1ηjΔzt−j+at, or
Δ y t = ϕ 10 + α 1 [ y t − 1 − β 0 x t − 1 ] + ∑ j = 1 p − 1 η 11 j Δ y t − j + ∑ j = 1 p − 1 η 12 j Δ x t − j + a 1 t , Δ x t = ϕ 20 + α 2 [ y t − 1 − β 0 x t − 1 ] + ∑ j = 1 p − 1 η 21 j Δ y t − j + ∑ j = 1 p − 1 η 22 j Δ x t − j + a 2 t , \begin{aligned} &\Delta y_{t}=\phi_{10}+\alpha_{1}\left[y_{t-1}-\beta_{0} x_{t-1}\right]+\sum_{j=1}^{p-1} \eta_{11}^{j} \Delta y_{t-j}+\sum_{j=1}^{p-1} \eta_{12}^{j} \Delta x_{t-j}+a_{1 t}, \\ &\Delta x_{t}=\phi_{20}+\alpha_{2}\left[y_{t-1}-\beta_{0} x_{t-1}\right]+\sum_{j=1}^{p-1} \eta_{21}^{j} \Delta y_{t-j}+\sum_{j=1}^{p-1} \eta_{22}^{j} \Delta x_{t-j}+a_{2 t}, \end{aligned} Δyt=ϕ10+α1[yt−1−β0xt−1]+j=1∑p−1η11jΔyt−j+j=1∑p−1η12jΔxt−j+a1t,Δxt=ϕ20+α2[yt−1−β0xt−1]+j=1∑p−1η21jΔyt−j+j=1∑p−1η22jΔxt−j+a2t,
where α 1 \alpha_{1} α1 and α 2 \alpha_{2} α2 are adjustment coefficients. -
For a cointegrated system, you can not apply VAR model directly to the first difference process.
-
Overdifferencing
-
e t − 1 = y t − 1 − β 0 x t − 1 e_{t-1}=y_{t-1}-\beta_{0} x_{t-1} et−1=yt−1−β0xt−1 is a “compensation” term for the overdifferenced system Δ z t \Delta z_{t} Δzt.
The Engle-Granger Methodology
-
Step 1: Pretest { x t } \left\{x_{t}\right\} {xt} and { y t } \left\{y_{t}\right\} {yt} for their order of integration.
-
Step 2 : If the results of Step 1 indicate that both { x t } \left\{x_{t}\right\} {xt} and { y t } \left\{y_{t}\right\} {yt} are I ( 1 ) \mathrm{I}(1) I(1), the next step is to estimate the long-run equilibrium relationship in the form:
y t = β 0 + β 1 x t + e t . y_{t}=\beta_{0}+\beta_{1} x_{t}+e_{t} . yt=β0+β1xt+et.
Save the residuals { e ^ t } \left\{\widehat{e}_{t}\right\} {e t}. -
Step 3: Consider the autoregression of the residuals:
Δ e ^ t = θ e ^ t − 1 + ∑ i = 1 q a i Δ e ^ t − i + u t \Delta \widehat{e}_{t}=\theta \widehat{e}_{t-1}+\sum_{i=1}^{q} a_{i} \Delta \widehat{e}_{t-i}+u_{t} Δe t=θe t−1+i=1∑qaiΔe t−i+ut
Test H 0 : θ = 0 H_{0}: \theta=0 H0:θ=0. Be aware that the critical values are different from those in the Dickey-Fuller tables. If we reject the null hypothesis θ = 0 \theta=0 θ=0, we can conclude that the residual sequence is stationary and that { x t } \left\{x_{t}\right\} {xt} and { y t } \left\{y_{t}\right\} {yt} are cointegrated. -
Step 4 : Estimate the error-correction model
Δ y t = ϕ 10 + α 1 e ^ t − 1 + ∑ j = 1 p − 1 η 11 j Δ y t − j + ∑ j = 1 p − 1 η 12 j Δ x t − j + a 1 t Δ x t = ϕ 20 + α 2 e ^ t − 1 + ∑ j = 1 p − 1 η 21 j Δ y t − j + ∑ j = 1 p − 1 η 22 j Δ x t − j + a 2 t \begin{aligned} &\Delta y_{t}=\phi_{10}+\alpha_{1} \widehat{e}_{t-1}+\sum_{j=1}^{p-1} \eta_{11}^{j} \Delta y_{t-j}+\sum_{j=1}^{p-1} \eta_{12}^{j} \Delta x_{t-j}+a_{1 t} \\ &\Delta x_{t}=\phi_{20}+\alpha_{2} \widehat{e}_{t-1}+\sum_{j=1}^{p-1} \eta_{21}^{j} \Delta y_{t-j}+\sum_{j=1}^{p-1} \eta_{22}^{j} \Delta x_{t-j}+a_{2 t} \end{aligned} Δyt=ϕ10+α1e t−1+j=1∑p−1η11jΔyt−j+j=1∑p−1η12jΔxt−j+a1tΔxt=ϕ20+α2e t−1+j=1∑p−1η21jΔyt−j+j=1∑p−1η22jΔxt−j+a2t -
Step 5: Assess model adequacy.
Johansen Methodology
- The Engle-Granger (1987) methodology seeks to determine whether the residuals of the equilibrium relationship are stationary.
- The Johansen (1988) methodology determines the rank of Π = Φ 1 + Φ 2 + ⋯ + Φ p − I k \Pi=\Phi_{1}+\Phi_{2}+\cdots+\Phi_{p}-I_{k} Π=Φ1+Φ2+⋯+Φp−Ik
- He developed a sophisticated sequential procedure for determining the existence of cointegration and for determining the number of cointegrating relationships based on maximum likelihood techniques.
The rank of Π = Φ 1 − I k \Pi=\Phi_{1}-I_{k} Π=Φ1−Ik equals the number of linearly independent cointegrating vectors.
- If Π = 0 \Pi=0 Π=0 - no cointegration! This system contains exactly k k k unit roots ( k k k stochastic trends).
- is singular but nonzero, say ran k ( Π ) = m < k \operatorname{ran} k(\Pi)=m<k rank(Π)=m<k. This system contains k − m k-m k−m unit roots that give k − m k-m k−m stochastic trends.
- If Π \Pi Π is of full rank-unity does not solve Eq.(10). Thus z t z_{t} zt is stationary (no stochastic trend).
有几个单位根说明有几个stochastic trend, 行列式的秩表明是几阶cointegration
Error Correction Model
-
For example, if m = 2 m=2 m=2, Eq. (9) becomes
Δ z 1 t = ϕ 10 + α 1 ( 1 ) e 1 t − 1 + α 2 ( 1 ) e 2 t − 1 + a 1 t Δ z 2 t = ϕ 20 + α 1 ( 2 ) e 1 t − 1 + α 2 ( 2 ) e 2 t − 1 + a 2 t ⋮ Δ z k t = ϕ k 0 + α 1 ( k ) e 1 t − 1 + α 2 ( k ) e 2 t − 1 + a k t \begin{gathered} \Delta z_{1 t}=\phi_{10}+\alpha_{1}^{(1)} e_{1 t-1}+\alpha_{2}^{(1)} e_{2 t-1}+a_{1 t} \\ \Delta z_{2 t}=\phi_{20}+\alpha_{1}^{(2)} e_{1 t-1}+\alpha_{2}^{(2)} e_{2 t-1}+a_{2 t} \\ \vdots \\ \Delta z_{k t}=\phi_{k 0}+\alpha_{1}^{(k)} e_{1 t-1}+\alpha_{2}^{(k)} e_{2 t-1}+a_{k t} \end{gathered} Δz1t=ϕ10+α1(1)e1t−1+α2(1)e2t−1+a1tΔz2t=ϕ20+α1(2)e1t−1+α2(2)e2t−1+a2t⋮Δzkt=ϕk0+α1(k)e1t−1+α2(k)e2t−1+akt -
Granger representation theorem applies in the more general k \mathrm{k} k-variable VAR ( p ) \operatorname{VAR}(\mathrm{p}) VAR(p) models for I ( 1 ) I(1) I(1) variables. If z t z_{t} zt is cointegrated, the error correction model can be similarly constructed.
-
The eigenvalues (with some normalization) denoted λ i \lambda_{i} λi are put in order:
λ 1 ≥ λ 2 ≥ ⋯ ≥ λ k − 1 ≥ λ k \lambda_{1} \geq \lambda_{2} \geq \cdots \geq \lambda_{k-1} \geq \lambda_{k} λ1≥λ2≥⋯≥λk−1≥λk -
λ i ∈ [ 0 , 1 ) \lambda_{i} \in[0,1) λi∈[0,1), for i = 1 , ⋯ , k i=1, \cdots, k i=1,⋯,k
-
If z t z_{t} zt are not cointegrated, then rank ( Π ) = 0 \operatorname{rank}(\Pi)=0 rank(Π)=0. Therefore, all λ i = 0 \lambda_{i}=0 λi=0. Consequently, log ( 1 − λ i ) = 0 \log \left(1-\lambda_{i}\right)=0 log(1−λi)=0 for all i i i.
-
If rank ( Π ) = 1 \operatorname{rank}(\Pi)=1 rank(Π)=1, then log ( 1 − λ 1 ) \log \left(1-\lambda_{1}\right) log(1−λ1) will be negative and log ( 1 − λ i ) = 0 \log \left(1-\lambda_{i}\right)=0 log(1−λi)=0 for i ≥ 2 i \geq 2 i≥2.
-
If Π \Pi Π is of full rank, then log ( 1 − λ i ) < 0 \log \left(1-\lambda_{i}\right)<0 log(1−λi)<0 for all i i i.
trace cointegration tests
A sequential procedure to determine the number of cointegrating vectors.
- L R trace ( m ) = − ( T − p ) ∑ i = m + 1 k log ( 1 − λ ^ i ) L R_{\text {trace }}(m)=-(T-p) \sum_{i=m+1}^{k} \log \left(1-\widehat{\lambda}_{i}\right) LRtrace (m)=−(T−p)∑i=m+1klog(1−λ i), where λ ^ i \widehat{\lambda}_{i} λ i is the estimated value for λ i \lambda_{i} λi from the Π ^ \widehat{\Pi} Π matrix.
- L R trace ( m ) L R_{\text {trace }}(m) LRtrace (m) is used to test H 0 : rank ( Π ) ≤ m H_{0}: \operatorname{rank}(\Pi) \leq m H0:rank(Π)≤m against H 1 : rank ( Π ) > m H_{1}: \operatorname{rank}(\Pi)>m H1:rank(Π)>m for m = 0 , 1 , 2 , ⋯ , k − 1 m=0,1,2, \cdots, k-1 m=0,1,2,⋯,k−1.
- Due to the presence of unit roots, the asymptotic distributions of the test statistics are nonnormal. The critical values depend on :
(1) the value of k − m k-m k−m, the number of unit roots
(2) the specification of μ t \mu_{t} μt
Johansen’s test : maximum eigenvalue tests
A sequential procedure to determine the number of cointegrating vectors.
- L R max ( m ) = − ( T − p ) log ( 1 − λ ^ m + 1 ) L R_{\max }(m)=-(T-p) \log \left(1-\widehat{\lambda}_{m+1}\right) LRmax(m)=−(T−p)log(1−λ m+1)
- L R max ( m ) L R_{\max }(m) LRmax(m) is used to test H 0 : rank ( Π ) = m H_{0}: \operatorname{rank}(\Pi)=m H0:rank(Π)=m against H 1 : rank ( Π ) = m + 1 H_{1}: \operatorname{rank}(\Pi)=m+1 H1:rank(Π)=m+1 for m = 0 , 1 , 2 , ⋯ , k − 1 m=0,1,2, \cdots, k-1 m=0,1,2,⋯,k−1.
- Critical values of the test statistics are nonstandard and depend on the specification of μ t \mu_{t} μt and the number of unit roots. And they must be evaluated via simulation.
Cointegration Test Specification
The limiting distributions of cointegration tests depend on the deterministic function
μ
t
\mu_{t}
μt.
-
Case 1 μ t = 0 1 \mu_{t}=0 1μt=0. In this case, all the component series of z t z_{t} zt are I ( 1 ) I(1) I(1) without drift and the stationary series e t = β ′ z t e_{t}=\beta^{\prime} z_{t} et=β′zt has mean zero.
-
Case 2 μ t = μ 0 = α c 0 2 \mu_{t}=\mu_{0}=\alpha c_{0} 2μt=μ0=αc0, where c 0 c_{0} c0 is an m-dimensional nonzero constant vector. The ECM becomes
Δ z t = α ( β ′ z t − 1 + c 0 ) + η 1 Δ z t − 1 + ⋯ + η p − 1 Δ z t − p + 1 + a t \Delta z_{t}=\alpha\left(\beta^{\prime} z_{t-1}+c_{0}\right)+\eta_{1} \Delta z_{t-1}+\cdots+\eta_{p-1} \Delta z_{t-p+1}+a_{t} Δzt=α(β′zt−1+c0)+η1Δzt−1+⋯+ηp−1Δzt−p+1+at
so that the components of z t z_{t} zt are I ( 1 ) I(1) I(1) without drift, but e t e_{t} et have a nonzero mean − c 0 -c_{0} −c0. This is referred to as the case of restricted constant. -
Case 3 μ t = μ 0 3 \mu_{t}=\mu_{0} 3μt=μ0, which is nonzero. Here the component series of z t z_{t} zt are I ( 1 ) I(1) I(1) with drift μ 0 \mu_{0} μ0 and e t e_{t} et may have a nonzero mean.
-
Case 4 (not common in empirical work) μ t = μ 0 + α c 1 t \mu_{t}=\mu_{0}+\alpha c_{1} t μt=μ0+αc1t, where c 1 c_{1} c1 is an m-dimensional nonzero constant vector. The ECM becomes
Δ z t = μ 0 + α ( β ′ z t − 1 + c 1 t ) + η 1 Δ z t − 1 + ⋯ + η p − 1 Δ z t − p + 1 + a t \Delta z_{t}=\mu_{0}+\alpha\left(\beta^{\prime} z_{t-1}+c_{1} t\right)+\eta_{1} \Delta z_{t-1}+\cdots+\eta_{p-1} \Delta z_{t-p+1}+a_{t} Δzt=μ0+α(β′zt−1+c1t)+η1Δzt−1+⋯+ηp−1Δzt−p+1+at
so that z t z_{t} zt is / ( 1 ) /(1) /(1) with drift μ 0 \mu_{0} μ0 and e t e_{t} et has a linear time trend related to c 1 t c_{1} t c1t. This is the case of restricted trend. -
Case 5 (not common in empirical work) μ t = μ 0 + μ 1 t \mu_{t}=\mu_{0}+\mu_{1} t μt=μ0+μ1t, where μ i \mu_{i} μi are nonzero. Here both the constant and trend are unrestricted. z t z_{t} zt is I ( 1 ) I(1) I(1) and has a quadratic time trend and e t e_{t} et has a linear trend.